yayaiu6/Verto-Arabic-ASR

Verto Arabic ASR

Egyptian Arabic and mixed Arabic–English speech recognition, packaged as ONNX INT8 for browser use. Audio is processed on your device.

Try the demo · Open in Colab · Code and deployment

Model

• Conformer CTC model. Verto provides the INT8 conversion and browser integration; it is not a newly trained model.
• 69.46 MB for the ONNX file; 69.93 MB including tokenizer and audio-processing assets. Sizes are decimal MB; browser runtime files are separate.
• Weight-only INT8 with FP32 activations; some weights remain FP32. Greedy CTC decoding, without KenLM or timestamps.
• The browser application accepts 0.5–60 seconds of audio. CPU/WASM is the default; WebGPU is optional with CPU fallback.

Use in JavaScript

Serve the complete application from the GitHub repository, then import its browser client:

import { VertoASR } from '/assets/app/verto.mjs';

const asr = new VertoASR();
try {
  await asr.load({ device: 'wasm' });
  const result = await asr.transcribe(file); // Browser File object
  console.log(result.text);
} finally {
  asr.close();
}

The app uses ONNX Runtime Web with Verto's audio preprocessing and text decoding. This model is not a drop-in Transformers.js pipeline. Docker deployment needs no Node or npm. Full setup is in the repository README.

For direct ONNX use: mono 16 kHz audio → 80-bin log-Mel features. See preprocessor.json and mel_filters.json. Inputs: features float32 [1, T, 80], feature_lengths int64 [1]. Outputs: log_probs float32 [1, T_out, V], output_lengths int64 [1]. The bundled tokenizer.json is the app's piece table, not Hugging Face Tokenizers format.

Evaluation

Compared with the original Metro FP32 checkpoint on 6,726 Casablanca recordings (7.70 hours), using the Open Universal Arabic ASR Leaderboard's evaluator and greedy CTC without KenLM. Lower error rates are better.

Metric	Original FP32	Verto INT8
WER — all eight dialects	75.408%	75.449%
CER — all eight dialects	41.847%	41.875%
WER — Egyptian Arabic (825 recordings)	30.678%	31.115%
Inference file size	238.84 MB	69.46 MB

The file is 70.92% smaller, with WER increases of 0.041 percentage points overall and 0.437 on Egyptian Arabic. Both versions have high error rates on several other dialects. These results do not establish equal accuracy on other audio; speed, memory and browser performance were not benchmarked.

Full benchmark and methodology · Results CSV

Status and source

Preview: evaluated against Casablanca reference transcripts; see the evaluation above. Accuracy varies with noise, dialect and unfamiliar words. No claim of unchanged accuracy versus the original model is made.

Source weights: Metro-ASR Space revision 525644dc782b6627f1eb1f0a589407fcacbfc152. The existing MIT license and original attribution are preserved in LICENSE.